All-sky assimilation of SSMIS humidity sounding channels over land within the ECMWF system

Transcription

1 EUMETSAT/ECMWF Fellowship Programme Research Report No. 38 All-sky assimilation of SSMIS humidity sounding channels over land within the ECMWF system Fabrizio Baordo and Alan J. Geer July 2015 To be submitted to Quarterly Journal of the Royal Meteorological Society

2 Series: EUMETSAT/ECMWF Fellowship Programme Research Reports A full list of ECMWF Publications can be found on our web site under: Contact: library@ecmwf.int c Copyright 2015 European Centre for Medium Range Weather Forecasts Shinfield Park, Reading, RG2 9AX, England Literary and scientific copyrights belong to ECMWF and are reserved in all countries. This publication is not to be reprinted or translated in whole or in part without the written permission of the Director-General. Appropriate non-commercial use will normally be granted under the condition that reference is made to ECMWF. The information within this publication is given in good faith and considered to be true, but ECMWF accepts no liability for error, omission and for loss or damage arising from its use.

3 Abstract The extension of all-sky assimilation of SSMIS humidity sounding channels to land surfaces is investigated in this paper. A symmetric error model, which adaptively determines the size of observation errors, can be formulated using the scattering index as a predictor to identify cloudy and precipitating regions over land. This assigns larger observation errors in those situations more difficult to model because of radiative transfer and mislocation errors. The use of an instantaneous emissivity retrieval from SSMIS surfacesensitive channels is also explored. In clear-sky scenes, emissivity retrievals appear better at capturing daily differences in surface conditions, compared to emissivity atlas values. In the presence of clouds, retrievals have different behaviour. In the lower microwave frequencies (less than 50 GHz), emissivity estimates appear nearly as reliable as those in clear-skies, but at higher frequencies, as the magnitude of scattering increases, so does the error in the retrieval and the resultant emissivity estimate can be unphysically low or high. However, the retrieval still appears feasible at high frequencies in light cloud situations; the number of retrievals discarded due to these kind of problems is around 10%. In these cases, an estimate from an emissivity atlas can be substituted instead. Assimilation experiments are performed that demonstrate the feasibility of assimilating SSMIS 183 GHz channels over land in all-sky conditions: the assimilation system is not degraded and the improvements on analysis and forecast scores are about the same as those which are obtained by the equivalent clear-sky approach. The developments described in this study were an essential first step to create framework to allow the all-sky assimilation over land of other microwave humidity sounders: this started operationally at ECMWF in 2015, covering both SSMIS and four MHS (Microwave Humidity Sounder) instruments. 1 Introduction The assimilation of microwave observations in numerical weather prediction (NWP) systems is still more intensive over ocean than over land surfaces. Over ocean, fast and accurate emissivity models have been developed for NWP (e.g. Liu et al., 2011; Kazumori and English, 2014) that allow the assimilation of channels with strong surface-sensitivity. In contrast, the microwave signal emerging from land surfaces not only depends on frequency, incidence angle and polarisation, but is also affected by the large variability in surface types (e.g deserts, vegetation, high orography) and conditions (e.g. roughness, moisture, snow, ice). For this reason, in data assimilation, the high complexity of modelling the interaction between all these surface parameters and the microwave radiation has generally restricted the use of observations to temperature and humidity sounding channels which receive a weak contribution from the surface. However, an estimate of surface emissivity is still required to assimilate such data. Increasingly in NWP, the approach is to make retrievals of land emissivities directly from satellite measurements (e.g. Karbou et al., 2005; Prigent et al., 2005; Ruston et al., 2008 among many others). The general assumption is that for most surface types, land emissivity is sufficiently invariant with frequency that emissivity retrieved from surface-sensitive ( window ) channels can be used as a reasonable approximation for sounding channels. This approach has also been implemented for some years within the operational ECMWF system to assimilate clear-sky observations in temperature and humidity sounding channels (Krzeminski et al., 2009) from AMSU-A (Advanced Microwave Sounding Sounding Unit-A) and MHS (Microwave Humidity Sounder): emissivities retrieved from AMSU-A channel 3 (50.3 GHz) are given to the temperature channels (50-60 GHz), whilst retrievals from MHS channel 1 (89 GHz) are assigned to the humidity channels (183 GHz). Karbou et al., 2010a,b investigated the use of land emissivity retrievals within the Météo France assimilation system, showing that it helps to reduce bias and standard deviation of first-guess (FG) departures (e.g. observation minus forecast differences) and increases the number of assimilated observations from AMSU-A and AMSU-B surface-sensitive and sounding channels. Improved correlations between observations and simulations were found over snow-covered areas, and analyses and forecasts over tropical regions appeared to be improved. Research Report No. 38 1

4 A further increase in the use of sounding observations over land can be achieved by extending the assimilation to clear, cloudy or precipitating scenes (the so called all-sky approach). However, at microwave frequencies in situations where the atmospheric scattering is most important (such as over land and in temperature and humidity sounding channels), the inaccuracy of scattering radiative transfer models has been a problem. (e.g. Geer et al., 2012; Baordo et al., 2012). However, Geer and Baordo (2014a) were able to globally improving the accuracy of the microwave scattering signal across all frequencies from 10 GHz to 183 GHz in all weather conditions by modelling snow as a non-spherical hydrometeor and uses optical properties from the Liu (2008) discrete dipole approximation for a sector snowflake. These developments allowed ECMWF to start operational all-sky assimilation of SSMIS (Special Sensor Microwave Imager Sounder) humidity sounding channels over ocean (Geer, 2013). This brings benefits to forecast quality; the all-sky assimilation of the SSMIS humidity channels brings roughly twice the benefit of just assimilating the observations in clear skies. The benefit comes through the 4D-Var assimilation, which can infer dynamical initial conditions from humidity, cloud and precipitation features in the observations. This has also been demonstrated in the assimilation of clear-sky infrared humidity observations by Peubey and McNally (2009). The benefit of SSMIS humidity channels is greatest in the southern midlatitudes, where the storm-tracks provide ideal conditions for model-tracing. The good results over ocean strongly encouraged feasibility studies to extend the assimilation of SSMIS humidity sounding channels to land-surfaces. Our purpose in this paper is to document the technical and scientific changes which were needed to implement the all-sky framework over land, summarising and extending the initial work carried out by Baordo et al., 2012, The over land framework has been developed and evaluated through the use of SSMIS observations, but it can be generalised for the assimilation of other humidity sounder sensors. A general overview of the all-sky assimilation at ECMWF is provided in Sect. 2. Methodology to implement the over land framework is described in Sect. 3. Results of assimilation experiments are provided in Sect General overview of the all-sky assimilation 2.1 Radiative transfer model ECMWF has been operationally assimilating microwave imager observations in all-sky conditions for over five years (Bauer et al., 2010; Geer et al., 2010; Geer and Bauer, 2011). The observation operator designed for assimilating microwave radiances in clear, cloudy and precipitating scenes is RTTOV-SCATT (Bauer et al., 2006), which uses the delta-eddington approximation (Joseph et al., 1976) to solve the radiative transfer equation including scattering. The bulk optical properties for cloud water, cloud ice and rain are pre-tabulated for each hydrometeor type as a function of temperature, frequency and water content. Cloud water, cloud ice and rain are modelled as spherical particles using Mie theory and a constant density: the first two hydrometeors use a gamma size distribution (e.g. Petty and Huang 2011), while a Marshall and Palmer (1948) size distribution is used for rain. Since the ECMWF operational cycle 40r1, snow has been modelled as a non-spherical hydrometeor which uses optical properties calculated from the discrete dipole approximation for a sector snowflake (Liu, 2008). To simulate bulk optical properties for the Liu sector shape, the tropical version of the Field (2007) size distribution is used. Ocean surface emissivity is computed by version 5 of FASTEM (Liu et al., 2011; Bormann et al., 2012). Over land, as will be documented in Sect. 3, we adopt same strategy as implemented in previous studies (e.g. Karbou et al., 2010a,b; Krzeminski et al., 2009): we firstly retrieve land emissivities from satellite observations in surface-sensitive channels and, secondly, assuming that the spectral variability of emissivity is minimal, we reassign these estimates to the closest channel higher in frequency in order to perform radiative transfer calcu- 2 Research Report No. 38

5 lations (for instance, emissivities retrieved in the SSMIS 91 GHz channel are applied to the humidity sounding channels). If the retrieval fails, we use emissivities provided by the TELSEM (Tool to Estimate Land Surface Emissivities at Microwave) atlas, which is based on a pre-calculated monthly-mean emissivity climatology derived from 10 years of SSMI observations (Aires et al., 2011). TELSEM atlas values are available within RTTOV and they provide emissivity estimates for all land surfaces between 19 and 100 GHz and for all angles and linear polarisations. The observation operator in the all-sky framework selects the nearest model profile to the observation (in time and space) and then runs RTTOV-SCATT. The hydrometeor inputs to RTTOV-SCATT are the vertical profiles of cloud water, cloud ice, total rain and total snow, plus the effective cloud fraction (C), which controls the computation of the all-sky simulated observation (T all sky ) which is a weighted average of brightness temperature from two independent sub-columns, one clear (T clr ) and one cloudy (T cld ). The effective cloud fraction provides the weight given to the two sub-columns: T all sky = (1 C)T clr +CT cld. (1) Over ocean surfaces, C is computed as a hydrometeor-weighted average of cloud, convective and large-scale precipitation fractions across all vertical levels, providing an approximate but computationally efficient solution to account for the effects of sub-grid variability in cloud and precipitation (Geer et al., 2009a,b, the C av approach). Over land surfaces, C is computed as the largest cloud fraction in the model profile (the C max approach.) This is essentially a tuning measure to compensate for a relative lack of deep convection over land in the model, as compared to over the ocean. The choice of the Liu sector snowflake to model scattering from snow hydrometeors was made based on getting the best fit between modelled and observed brightness temperatures over ocean (Geer and Baordo, 2014a). It was not possible to find one particle that gave good results simultaneously over land and ocean. Instead of having different particle models for land and ocean, technically it was easier to vary the cloud fraction. With this in place, the accuracy of the cloud and precipitation radiative transfer model is as good over land as over ocean. 2.2 Observations In order to investigate the all-sky assimilation over land of 183 GHz humidity sounding channels, this study uses observations from SSMIS. In line with the ECMWF operational usage, only Defence Meteorological Satellite Program satellite (DMSP) F-17 has been used. The SSMIS sensor (Kunkee et al., 2008) represents an important advancement over its predecessor, SSMI, as it combines the SSMI imaging capabilities with the profiling capabilities of microwave sounders. SSMIS allows microwave measurements at frequencies ranging from 19 to 183 GHz over a swath width of 1707 km. SSMIS channels can be grouped as follows: 13 channels (channels 1-7 and 19-24) located near the oxygen absorption band (50-60 GHz) which allow atmospheric temperature sensing from about 80 km down to the earth surface; 3 humidity sounding channels (channels 9-11) close to the strong 183 GHz water vapour line; 7 SSMI-like imaging channels (channels 12-18), with the GHz replacing the SSMI 85.5 GHz channels and the addition of 1 channel at 150 GHz (channel 8). Table 1 summaries the general characteristics of SSMIS channels. The SSMIS F-17 data are pre-processed along the lines of Bell et al. (2008), who described the necessary corrections for F-16 observations. However, instrument improvements implemented for F-17 make the pre-processing slightly different: corrections are made for scan non-uniformity and reflector emission (based directly on an accurate thermistor-measured reflector temperature) but not for warm load intrusions, which are infrequent. In the all-sky framework, SSMIS observations are also averaged (or superobbed ) in boxes of approximately 80 Research Report No. 38 3

6 Table 1: Properties of SSMIS channels. Channel/NameFrequency PolarisationSensitivity (GHz) 1 / 50H 50.3 H Surface 2 / H Temperature 3 / H Temperature 4 / H Temperature 5 / H Temperature 6 / RC Temperature 7 / RC Temperature 8 / 150H H Humidity 9 / 183± ±6.6 H Humidity 10 / 183± ±3.0 H Humidity 11 / 183± ±1.0 H Humidity 12 / 19H H Surface 13 / 19V V Surface 14 / V Surface 15 / 37H 37.0 H Surface 16 / 37V 37.0 V Surface 17 / 91V V Surface 18 / 91H H Surface 19 / ± RC Temperature 20 / - ν = 60.79± RC Temperature / - ν ± RC Temperature 22 / - ν ± RC Temperature 23 / - ν ± RC Temperature 24 / - ν ± RC Temperature km by 80 km in order to make the horizontal scales of observed cloud and precipitation more similar to their effective resolution in the model (Geer and Bauer, 2010). 3 All-sky framework over land 3.1 Observation error modelling In the all-sky assimilation, the accuracy of the radiative transfer calculations in cloud and precipitating areas is not the only parameter to deal with. Forecast models are also affected by mislocation errors which are due to difficulties in predicting cloud and precipitation in exactly the right place with the right intensity (e.g. Fabry and Sun, 2010). Radiative transfer error and mislocation error lead to a highly non-gaussian behaviour of first-guess departures, independent of the surface type (land or ocean). Geer and Bauer (2011) showed that this representivity problem can be resolved by means of a symmetric error model which can provide a robust threshold quality-control check and also determine the size of observation errors for the assimilation. For microwave imager observations over ocean, the observation error is a linear function of the symmetric cloud amount given by the average of observed and simulated polarisation difference at 37 GHz. 4 Research Report No. 38

7 a Mean [kg/m 2 ] LHC+IHC IHC Observed SI [91H 150H] b 4 Mean [kg/m 2 ] Simulated SI [91H 150H] Figure 1: Mean of total (solid line) and ice (dashed-dotted line) hydrometeor content (kg m 2 ) binned as a function of the observed (a) and the simulated (b) scattering index. The liquid hydrometeor content (LHC) is given by the sum of model FG cloud water and rain, while the ice hydrometeor content (IHC) takes into account cloud ice and snow. Sample of SSMIS data is for the month of June 2013 considering observations over land restricted to latitudes equatorward of 60. Bin size is 0.5 K. Over land, the 37 GHz polarisation difference cannot be used as a cloud predictor for the symmetric error model, as it relies on the highly polarised nature of ocean surface emission. Instead, the greatest atmospheric signal from hydrometeors comes in the higher frequencies from snow, so the scattering index (SI, e.g. Bennartz et al., 2002; Bennartz and Bauer, 2003), is a good parameter to identify precipitation and convective areas. A SI given by the difference between SSMIS channel 18 (91H) and channel 8 (150H) was found to give good performance as a symmetric predictor for the observation error model. (Baordo et al., 2012). To check the capability of the SI to identify clear-sky areas as well as regions affected by clouds and precipitation, we can compare the modelled SI to the total hydrometeor content in the model, given by the sum of the model FG cloud water, cloud ice, rain and snow. We can also distinguish between the liquid hydrometeor content (LHC), considering only cloud water and rain, and the ice hydrometeor content (IHC), taking into account only cloud ice and snow. Fig. 1b shows that the model s mean hydrometeor content is approximately a linear function of the simulated SI. Cloud-free regions (zero hydrometeor content) are characterised by SI less than 0 K, while for increasing values of SI, the hydrometeor content increases as well, representing the increasing influence of scattering as the modelled cloud and precipitation amount increases. The IHC and the total hydrometeor content are relatively similar, suggesting that the LHC adds little information to the main signal, which comes from cloud ice and snow and appears to be well identified by the SI. There is not such an obvious link between the observed SI and the IHC (Fig. 1a). This is the consequence of mislocation errors. There are some cases where the observation is cloudy (higher scattering index), but there is no corresponding cloud in the model, counterbalanced by other cases where the observation is cloud-free (lower scattering index), but the model has cloud. Research Report No. 38 5

8 Error [K] SI sym [91H 150H] Figure 2: Error model for SSMIS channel 9 (183±7 GHz), showing how the standard deviations of FG departures binned as a function of the symmetric scattering index (solid line) are modelled by a linear fit (dotted line). FG departures are for the month of June 2013 considering SSMIS observations over land restricted to latitudes equatorward of 60. Bin size is 0.5 K. We can now derive the observation error formulation from the symmetric scattering index SI sym, which is the mean of the observed and the simulated SI. Fig. 2 gives an example of how the standard deviations of SSMIS channel 9 FG departures have been binned as a function of SI sym ; results for SSMIS channel 10 and 11 are similar in shape if not magnitude (not shown). For low scattering indexes, i.e. the clear-sky regime, the standard deviation is relatively constant at around 3 K. As the symmetric mean SI increases, so does the standard deviation. This increase is reasonably well-modelled by a straight line. However, beyond a certain threshold the errors seem to saturate. This behaviour is modelled by a piece-wise linear fit that is estimated graphically and shown on the figure. For this channel the model has a minimum clear-sky standard deviation E min of 3 K and a maximum E max of 57 K in strongly scattering situations. The start and end points of the up-slope are labelled SI clr and SI cld, with standard deviation of FG departures T err being predicted as follows: T err = E min SI sym SI clr ; (2) T err = E max + (E max E min ) SI sym SI clr SI cld SI clr SI clr < SI sym < SI cld ; (3) T err = E max SI sym SI cld. (4) The standard deviation of the FG departures is a combination of background error and observation error, so to convert T err to a predicted observation error, we subtract (in quadrature) an estimated background error. This is 1 K for all channels, converting a clear sky FG departure standard deviation of 3 K into an observation error of 2.8 K. This is a little larger than the constant observation error of 2 K which has been used to assimilate MHS humidity sounding observations over both land and ocean in clear-sky. The slightly larger observation error reflects the standard deviations in situations where SI is around 0 K; it is a cautious approach to assimilating all-sky SSMIS data over land, even when nominally in clear skies. The effectiveness of the observation error is discussed later in Sect Research Report No. 38

9 3.2 Surface emissivity retrieval For a non scattering plane-parallel atmosphere, assuming a flat and specular surface, for a given zenith (θ) angle and frequency (ν), the brightness temperature (T b ) observed by a satellite sensor can be expressed as follows: T b = ε (θ,ν) T s Γ (θ,ν) + (1 ε (θ,ν) )T (θ,ν) Γ (θ,ν) + T (θ,ν), (5) where ε (θ,ν) represent the surface emissivity at observation zenith angle θ and frequency ν; T s, T (θ,ν) and T (θ,ν) are, respectively, the surface skin temperature, the atmospheric down-welling radiation at the surface and upwelling radiation at the top of the atmosphere; Γ (θ,ν) is the net atmospheric transmissivity. The emissivity can be retrieved from equation 5 (for simplicity we omit the dependency from zenith angle and frequency): ε = T b (T + T Γ) (T s T. (6) )Γ Eq. 6 represents the scheme that has been commonly used in literature to retrieve emissivity directly from satellite measurements for those channels which receive a strong contribution from the surface (e.g. Karbou et al., 2005; Prigent et al., 2005). In the case of NWP systems, the atmospheric contribution to the observed brightness temperature (T, T and Γ) is computed within the radiative transfer model using the atmospheric profiles from short-range forecasts as input. The surface skin temperature, like emissivity, is also affected by the land surface variability (e.g. soil-moisture, roughness, wetness, snow) so that it is a difficult parameter to estimate accurately (English, 2008). Skin temperature can be also retrieved from satellite measurements (e.g. Karbou et al., 2010a) and used within NWP systems. In our study, we did not investigate the impact of using different sources of skin temperature (e.g. retrieved or monthly-mean based on climatology), but following the existing clear-sky approach at ECMWF, we relied on the estimate provided by the short-range forecasts. In order to be consistent with the way the observations are used in the all-sky assimilation, we take account of two independent columns, one clear and one cloudy, which, weighted by the effective cloud fraction C, give the simulated brightness temperature (Eq. 1). Assuming that skin temperature and emissivity are the same in each sub-column, from Eq. 1 and Eq. 5, we have: ε = T b (1 C)(T clr + T clr Γ clr) C(T cld + T (1 C)(T s T clr )Γ clr +C(T s T cld )Γ cld cld Γ cld). (7) The atmospheric terms are simulated for the clear and cloudy sub-column (T clr, T clr, Γ clr and T cld, T cld, Γ cld). When the contribution of the cloudy column is zero (C = 0), the emissivity computation is reduced to the emissivity scheme of Eq. 6. Independently of the way the emissivity retrieval is done (Eq. 7 or Eq. 6) it is not free from error. The larger the error in model skin temperature or atmospheric profile, the bigger the uncertainty on the emissivity estimate. The assumption of specular surface can also generate errors in those cases where the validity of such approximation might be no longer valid (e.g. in presence of very rough terrain and surface scattering). Finally, where there is strong atmospheric scattering or high liquid water content, the increasing opacity of the atmosphere can amplify the magnitude of the error, to the point where the surface becomes invisible and an emissivity retrieval Research Report No. 38 7

10 is clearly impossible. Because the all-sky assimilation does not implement a cloud screening and observed brightness temperatures can be from clear, cloudy or precipitating scenes, this can affect our retrievals; later sections will examine the errors caused by cloud and precipitation and ways of dealing with them. Essentially, we default to an emissivity atlas value where the emissivity retrieval appears to be erroneous. 3.3 Emissivity retrieval assessment The technical implementation of the emissivity retrieval in the all-sky framework was first evaluated by Baordo et al. (2012) but here we will examine the impact of cloud in more detail and look at the benefit of using instantaneous emissivity estimates rather than values provided by climatology. This is organised as follows: In Sect , we look at emissivity retrievals obtained using cloud free observations. Secondly, in Sect , we examine the impact of cloud on the retrievals. Finally, in Sect , we describe the the completed scheme for generating emissivities to be used for the 183 GHz channels. To compare the clear-sky and all-sky emissivity retrievals outlined in the previous section, we ran two assimilation experiments (further details are provided in section 4). Both process SSMIS observations through the all-sky path of the ECMWF system but the first experiment implements Eq. 7, the second, forcing C = 0, retrieves emissivity through Eq. 6. Even though the two experiments have slightly different backgrounds (e.g. different skin temperature and atmospheric profile coming from the natural chaotic variability that occurs when two non-identical assimilation experiments are performed) it is still reasonable to compare the two calculations of surface emissivity as they use identical SSMIS observations. Retrievals are performed globally but two screening criteria are applied. First, retrievals are restricted to latitudes equatorward of 60, particularly to avoid high-latitude areas where, during winter, snow and ice cover might introduce additional errors in the emissivity retrieval. Second, SSMIS observations are rejected in those locations where the model grid-point contains a mixture of water and land. Retrievals are computed only if the model land-sea-mask is greater than It is worth mentioning that in this study, before the emissivity computation (Eq. 6 or Eq. 7), bias corrections are applied to the satellite observations. In the ECMWF system, for microwave instruments, biases are inferred as a function of predictors including scan angle, the surface wind speed and the layer thickness. Bias coefficients are derived within the analysis system using variational bias correction (Auligné et al., 2007). In the following sections, we also refer to Eq. 6 and Eq. 7, respectively, as the clear-sky (ε cs ) and all-sky (ε as ) emissivity retrieval Retrieval in cloud free scenes As an example of clear-sky retrievals we chose a bare soil point in North Africa. The location is in Algeria (28.42 N;4.50 E) and it was observed by 44 SSMIS overpasses for the period ranging from 1 (5 UTC) to 30 (17 UTC) June 2013, shown in Fig. 3. The observations are most likely free from clouds and precipitation during the entire month: the SSMIS observed and simulated SI are always less than 0; also images from SEVIRI show no cloud systems over the selected location. The modelled atmospheric transmittance is always high (panel d) so the retrievals should be reliable. The retrieved emissivities are shown for channels 19H and 91H, alongside the value from the TELSEM atlas (panels a and b). Because the model cloud fraction is always nearly zero at this location, there is little difference between retrievals from the all-sky or the clearsky approach. Retrieved emissivity is similar to the atlas but shows variability from one overpass to the next. Particularly at 19H it appears to be anticorrelated with the diurnal variability of skin temperature (correlation coefficient is -0.90). Skin temperature varies around 15 K between the times of the early evening and early morning orbits (panel c). In dry desert soils the microwave emission may come from some distance below the 8 Research Report No. 38

11 a Emissivity H Number of SSMIS overpass b 91H Emissivity Number of SSMIS overpass c Temperature [K] Number of SSMIS overpass d Transmittance H 91H Number of SSMIS overpass Figure 3: All-sky emissivity retrievals derived by SSMIS measurements in the Northern of Africa (28.42 N;4.50 E) from 1 to 30 June 2013 (for a total of 44 overpasses): a) for 19 GHz horizontal polarisation b) for 91 GHz horizontal polarisation. The black thick line represents the TELSEM atlas value. To complete the figure, for each SSMIS overpass: c) model skin temperature; d) model transmittance at 19 and 91 GHz. Research Report No. 38 9

12 Table 2: RMSE of differences between simulated and SSMIS observed brightness temperatures in the Northern of Africa (28.42 N;4.50 E) from 1 to 30 June 2013 (for a total of 44 overpasses). T b as and T b cs indicate the simulations computed respectively using ε as and ε cs. RMSE [Tb as -Tb] RMSE [Tb cs -Tb] 183± ± ± surface. The anticorrelation is consistent with an excessively strong diurnal cycle of skin temperature in the model, perhaps because the diurnal cycle of sub-soil temperature is damped compared to the skin temperature (the forecast model skin temperature is a key part of radiative energy transfer at the surface, so it is more representative of infrared and visible wavelengths where it is a genuine skin temperature). Hence one benefit of doing an emissivity retrieval, rather than using an atlas, is that it reduces the impact of errors in the use of model skin temperature, which is particularly evident in desert regions. Although it would be nice to estimate a skin temperature simultaneously with the emissivity, that is not feasible (see Eq. 6). As a general consideration, the variability of our retrievals respect to TELSEM estimates might also be partially driven by the superobbed observations: the exact location of raw SSMIS observations composing the superob is going to vary from day to day and so we might see a slightly different bit of surface each time with a slightly different emissivity. Further, the 91H emissivity retrievals in Fig. 3 are lower than the equivalent atlas value. This is observed more generally, and might be explained considering that TELSEM estimates were derived from SSMI data at 85.5 GHz, and there may be some change in the emissivity with frequency. So, even in a desert environment where surface emissivity is normally quite constant with time, there are some advantages of using emissivity retrievals rather than an atlas: compensating for skin temperature errors; matching the emissivity of the exact field of view and other sensor characteristics (such as polarisation); reducing the impact of frequency extrapolation. The RMSE of differences between simulated and observed brightness temperatures can provide a general measure of the error in the radiative transfer calculations. Table 2, for the Algeria case, provides the RMSE computed considering the 44 SSMIS observations in the 183 GHz humidity sounding channels, which are assigned the emissivity retrieved at 91 GHz. We distinguish between RMSE computed using ε as to simulate the brightness temperatures (T b as ) and that coming from the use of ε cs (T b cs ), but results are very similar for both. The error is largest at 183±7 GHz in SSMIS channel 9. This is a lower tropospheric humidity channel and, consequently, in the relatively clear atmosphere of this example, it is likely to be more affected by errors in the surface parameters (e.g. emissivity, skin temperature). In any case, the residual uncertainty in the first guess is consistent with the observation error model, which in this scenario would predict a total error of 3 K. All 44 SSMIS humidity sounding observations in this example are actively assimilated and not rejected by quality control. To generalise the behaviour of the retrieval for SSMIS window channels in cloud free scenes (observed SI 0 K), Fig. 4 compares global mean and standard deviation of emissivities from ε as, ε cs and TELSEM for the month of June As mentioned, statistics are computed between 60 S and 60 N, only considering land-sea-mask values greater than The sample has been restricted slightly further by (a) only considering retrievals that have generated emissivities between 0 and 1 and (b) requiring the availability of TELSEM atlas values. The all-sky and clear-sky retrievals behave the same across all the frequencies and match the climatology values fairly well. Statistics at 91H are those particularly relevant for the goal of our assimilation and they show the negative bias that was highlighted in the single observation case (Fig. 3). 10 Research Report No. 38

13 1.0 Emissivity H 19V 37H 37V 50H 91H 91V Figure 4: Mean of land surface emissivity retrievals in SSMIS window channels for the month of June ε as (thin line) and ε cs (dashed line) retrieved from SSMIS cloud free observations are compared to TELSEM atlas values (thick line). Retrievals are restricted to latitudes equatorward of 60 and model land-sea-mask values greater than Error bars represent the standard deviation. Table 3: Impact of cloud contamination on ε as and ε cs for 3 observation cases over the north-east of Bolivia (15.09 S;63.28 W). The observed brightness temperatures (Tb) and the model surface to space transmittance (Γ clr ) are also shown for the SSMIS window channels. Emissivity from TELSEM atlas (ε t ) are provided as additional reference. 19H 19V 37H 37V 50H 91H 91V ε t Intense scattering (observed SI 34 K) Tb [K] Γ clr ε as ε cs Intermediate scattering (observed SI 7 K) Tb [K] Γ clr ε as ε cs Moderate scattering (observed SI 1.2 K) Tb [K] Γ clr ε as ε cs Research Report No

14 3.3.2 Retrieval in cloud-affected scenes At higher microwave frequencies, the presence of clouds and precipitation can substantially reduce the surface to space transmittance, leading to larger errors in the emissivity estimate. However, in the case of limited scattering and reasonably high surface to space transmittance, an instantaneous emissivity retrieval might still be possible. To illustrate the impact of cloud contamination on the emissivity retrievals, we chose a densely vegetated area in the north-east of Bolivia (15.09 S;63.28 W). Compared to the desert area examined earlier, this area exhibits higher emissivities and is often affected by deep convection. Looking at the SSMIS observed and simulated SI, and cross-checking with images from GOES (Geostationary Operational Environmental Satellites), we chose 3 scenes which are cloud contaminated but are characterised by different amounts of scattering. It is important to remember that the all-sky assimilation is affected by mislocation errors so what it is seen in the observations it might or might not be present in the model with the same intensity. The 3 cases explore the situation when the observations have more cloud than the model, or about the same: 1. Intense scattering (30 June UTC). Observations have much more cloud than model: observed SI 34 K; simulated SI 7.5 K; model total hydrometeor content 0.47 kg m 2 ; model ice hydrometeor content 0.39 kg m 2 ; model TCWV 52.4 kg m 2 ; model rain rate mm h Intermediate scattering (24 June UTC). Observations have little bit more cloud than model: observed SI 7 K; simulated SI 3.5 K; model total hydrometeor content 0.35 kg m 2 ; model ice hydrometeor content 0.15 kg m 2 ; model TCWV 51 kg m 2 ; model rain rate mm h Moderate scattering (25 June UTC). Observations and model have same amount of cloud: observed SI 1.2 K; simulated SI 1.2 K; model total hydrometeor content 0.26 kg m 2 ; model ice hydrometeor content 0.04 kg m 2 ; model TCWV 48.5 kg m 2 ; model rain rate mm h 1. Note that even in the moderate scattering case this is far from a clear-sky location: there is cloud and precipitation present. Table 3 shows, for each observation case, ε as and ε cs retrievals, model surface to space transmittance and observed brightness temperatures for the SSMIS window channels. The magnitude of the scattering signal can be clearly distinguished within the 3 cases: moving from 19 to 91 GHz, the higher is the frequency, the larger is the depression on the observation. In the intense scattering case, observed brightness temperatures at 37, 50 and 91 GHz are about 10, 20 and 30 K colder than those at 19 GHz. In the second case, the size of such differences is reduced by a half and, in the moderate scattering situation, the impact of cloud is hard to identify. In general the lower microwave frequencies (19 and 37 GHz) are not affected by cloud contamination and, independently of the magnitude of the scattering, the emissivity retrievals appear to be consistent with those provided by TELSEM atlas. However, at 50 GHz and above, the depressed brightness temperatures in the intermediate and intense scattering cases lead to unphysically low emissivity retrievals. For example, ε as = in the 91H channel. This problem affects the clear-sky and all-sky emissivity retrievals equally. Even in the all-sky case, these are situations where the model has substantially less scattering than the observations and will struggle to generate a good retrieval. The method of emissivity retrieval assumes that the modelled atmospheric parameters that go into the retrieval are correct, but that assumption is violated in the intense and intermediate scattering cases. However, the moderate scattering case demonstrates that instantaneous emissivity retrievals are still feasible. In this particular case, the emissivity retrieval at 91 GHz is similar to the atlas values. The retrieved emissivity at 91 GHz can be given to the 183 GHz channels, leading first guess departures for SSMIS channel 9, 10 and 11 that are respectively K, K and K, indicating a good fit between model and observations. This shows that the retrieval is sufficiently good for our purpose of assimilating 183 GHz channels. Note that in all three cases, at higher microwave frequencies, the modelled clear-sky sensitivity to the surface is reduced (for instance the value of the surface to space transmittance at 91H 12 Research Report No. 38

15 Emissivity H 19V 37H 37V 50H 91H 91V Figure 5: As Fig. 4, but considering only SSMIS cloudy affected observations. is roughly half that at 19H) but that does not seem itself to be the cause of problems in retrieving emissivity; rather it is the strong depression of observed brightness temperatures in the intense and moderate scattering cases. Figure 5 generalises the impact of cloud contamination on SSMIS emissivity retrievals by showing global mean and standard deviation of emissivities from ε as, ε cs and TELSEM for the month of June 2013, selecting only those cases where observed SI >0 K, indicating scattering. Retrievals at frequencies less than 50 GHz are consistent with the climatology estimates, but at 50 and 91 GHz, as a result of cloud contamination, means and standard deviations of ε as and ε cs often diverge from those computed using TELSEM values, generally becoming too low. An issue unique to the all-sky approach is that the presence of cloud in the model profiles can also cause problems. This is the opposite problem to the one just illustrated; here the model has much more cloud than is present in the observation. In this situation in Eq. 7, the weight of the cloudy column in the emissivity calculation is very high (C close to or equal 1) and Γ cld is relatively small (much smaller than Γ clr ). This leads to a small denominator in Eq. 7 and, if the observed brightness temperature is not consistent with the cloud and precipitation in the model, this generates unphysically high emissivities, often larger than 1. Once again, the number of these failures increases at higher frequencies where scattering is the dominant effect and Γ cld is much smaller than Γ clr. Overall this means that the all-sky method can generate more poor-quality retrievals in strongly-scattering conditions than the clear-sky method, because there are roughly double the opportunities for problems: if there is intense scattering in either the model or the observations. However, if we want to push the methodology into cloud and precipitation-affected regimes, it still seems desirable to do a retrieval that is consistent with the radiative transfer model that is being used. However, clearly we need to be able to identify those situations where an emissivity retrieval is still feasible and to discard those retrievals affected by large errors Surface emissivity for the 183 GHz channels It appears to be useful to do an instantaneous emissivity retrieval in order to capture local changes in surface conditions that can affect the emissivity. We have chosen to use the all-sky emissivity retrieval, rather than the clear-sky one, for consistency with the way the simulations are done and with a view to the future. However, to eliminate poor-quality retrievals, a number of screening stages have been applied. First, we decided to reject observations where the value of the symmetric scattering index is greater than 20 K. As a reference, for 1 month of SSMIS data, this quality control discards roughly 1.2% of observations. We are avoiding attempting Research Report No

16 to assimilate profiles where scattering is most intense, either in the model or the observation. Second, we chose to reject emissivity retrievals (but still potentially use the observations) where they are outside of the range expected physical behaviour, i.e. outside of the range 0.55 to 1.0. Third, we compared the retrieval to the emissivity atlas and rejected retrievals that were more than a certain threshold away from the atlas value. These thresholds have been estimated from the standard deviation of difference considering retrievals obtained only from cloudy-free observations (e.g. observed SI 0 K like those of Fig. 4) and the corresponding atlas values, calculated globally in every frequency in summer and winter time (e.g. June and January). An average of the two seasons standard deviations is used and can be labelled σ. The final threshold was chosen as roughly 2 times σ in order to increase the chances to use an emissivity retrieval rather than the atlas. Thresholds used in every frequency are as follows: 0.04 (19H), 0.03 (19V), 0.04 (37H), 0.03 (37V), 0.06 (50H), 0.09 (91H), 0.07 (91V). The final sample of emissivities at 91H which has been used to simulate the brightness temperatures for the 183 GHz channels is most relevant to our assimilation system. Mean land emissivity maps at 91H for the month of June 2013 are shown in Fig. 6 which compares the retrieved emissivities (but where the 10% failing quality checks have been replaced by TELSEM atlas values) to the estimates provided by the climatology. The statistics are based on a total of SSMIS observations. The general offset between the retrievals and the atlas at 91 GHz is clearly visible. However, the retrieval does globally a good job in capturing the main features which characterise different surface types (e.g. desert, high orography, densely vegetated areas). Figure 7 gives a final illustration of the strengths and weaknesses of emissivity retrieval methods in challenging conditions. It shows all-sky and clear-sky surface emissivity retrievals during the monsoon season over an area on the border between Bangladesh and India where land usage is dominated by rice cultivation. The observed SI (panel a) is regularly larger than 5 K, indicating frequent episodes of precipitation. The simulated SI is not shown; though it is often in reasonable agreement with the observed SI, it can become unreliable in the presence of strong convection, because to simulate 91 GHz and 150 GHz brightness temperatures requires use of emissivity estimates (and we have chosen to use retrievals at 37 GHz and 91 GHz respectively) that themselves can become erroneous through the presence of scattering. The 19 GHz emissivity retrievals are largely unaffected by scattering and clear-sky and all-sky retrievals are generally very similar through the period (panel b). Retrieved emissivity peaks at around 0.9 in late June and declines rapidly to 0.7 in mid-july; we could speculate that this is associated with irrigation of the rice fields. The atlas emissivities show a similar pattern while missing the shorter-term variability. There are a few cases, particularly in mid-august, where the all-sky retrieval generates unphysically low emissivity retrievals and the clear-sky retrieval is OK: these are likely to do with the presence of heavy precipitation in the model that is not seen in the observations. At 91 GHz (panel c) there is a larger scatter in the emissivity estimates from both clear-sky and all-sky and 2% of clear-sky and 27% of all-sky emissivity retrievals are outside the physical bounds (0-1) and are not shown on the figure. However, the majority of retrievals are around , similar to the 19 GHz retrievals and much higher than the atlas emissivity of It is possible that the atlas itself is suffering from cloud contamination here. Panel c shows the estimated emissivities given to the all-sky simulations at 183 GHz: because the quality check is based on atlas values, the retrieved emissivity is almost always rejected and atlas is used instead, even if it likely has problems in this location. However, the emissivity estimate is almost immaterial for 183 GHz assimilation here: throughout the period the clear-sky surface-to-space transmittance in the 183±7 GHz channel is around Overall, this example illustrates that atlas, clear-sky and all-sky retrievals can all have their problems, but at least for 183 GHz assimilation, the worst cases for emissivity retrieval are those where either the surface visibility is irrelevant anyway, or the strongly convective situations will be given a large observation error through the symmetric error model. In these cases, using an emissivity retrieval or a value from climatology is a secondary problem to just achieving the assimilation of strongly scattering-affected humidity sounding channels. 14 Research Report No. 38

17 a b Figure 6: Mean land emissivity maps at 91H for the month of June 2013 for the mix generated retrieval/atlas emissivities (b) and the corresponding TELSEM atlas values (a). Grid spacing is Assimilation experiments The all-sky microwave brightness temperatures are used alongside many other satellite and conventional observations within a 12-hour assimilation window in the operational ECMWF 4D-Var assimilation system (Rabier et al., 2000) which produces global analyses and forecasts. The observation minus model differences (the so called first guess departures), which drive the data assimilation system, are computed through the 12 h assimilation window using the non-linear forecast model at the highest available resolution to propagate the background atmospheric state forward in time. The incremental 4D-Var, at lower resolution (about 80 km), finds the 12 h forecast evolution that optimally fits the available observations using a linearised forecast model. The atmospheric control variables, which consist of transforms of surface pressure, humidity, temperature and the two horizontal wind components, are adjusted during the minimisation. Cloud and precipitation are not part of the control vector, but during the minimisation, they are diagnosed from the dynamical and humidity fields every time step. Hence, by adjusting the temperature and moisture profile at the observation location, it is possible to modify cloud and precipitation to allow the analysis to fit the all-sky observations. The next sections examine results from the all-sky and clear-sky assimilation of SSMIS humidity sounding Research Report No

18 a) SI Jun 2013 Jul 2013 Aug b) 19H emissivity Jun 2013 Jul 2013 Aug c) 91H emissivity Jun 2013 Jul 2013 Aug d) 187 +/ 7 emissivity Jun 2013 Jul 2013 Aug 2013 Figure 7: Time-series of SSMIS emissivities and scattering index at N, 88.5 E, which is a rice growing region on the India-Bangladesh border: a) Observed scattering index (91H-150H); b) Atlas (dashed line), clear-sky emissivity retrieval (empty circle) and all-sky emissivity retrieval (filled circle) at 19H; c) As b but at 91H; d) The emissivity used for assimilation at 183 GHz. 16 Research Report No. 38